zhengxi-views: a traceable fund-manager Agent Skill for Zheng Xi's published views

Why a manager's own sentences beat a model's recollection

Ask a general assistant what a Chinese fund manager thinks about optical communications and you will usually get a fluent paragraph with no year, no publication and no quotation. The README opens on exactly this failure: answers that sound informed but cannot be traced to anything the manager said. zhengxi-views is built for people who need the citation, not the summary.

The subject is Zheng Xi of E Fund (易方达), described in the repository as deputy general manager of the equity investment management department and a portfolio manager. The skill is aimed at analysts, fund-of-funds researchers and individual investors who already follow him and want to check a claim quickly, plus anyone studying how a discretionary growth manager's stated framework lines up with the positions he actually filed. It is not a screener for the whole market, and it is not a recommendation engine. The README repeats the boundary: research and study aid only, not investment advice.

Three layers: raw corpus, a distilled method, and real fund data

The repository is organised as an Agent Skill, meaning a SKILL.md file with YAML frontmatter declaring name and description, plus a references/ directory loaded progressively and a scripts/ directory. The three data layers are the substance.

The first is references/corpus/, described as Zheng Xi's complete public output from 2012 to 2026: the investment-operation sections of periodic reports (quarterly, semi-annual and annual), manager notes, media interviews, plus a biography and lists of funds currently and formerly managed. The second is references/method.md, a method framework distilled from that corpus, where the README states every item carries a quotation from Zheng Xi himself. The third is references/fund_data/, real snapshots for all eight of his funds, four current and four former, each with quarterly top-ten holdings and a file covering net value, performance, size, asset allocation and tenure return. A separate references/all_funds/fund_list.json holds a market-wide list of roughly 27,000 funds that the fetch scripts can pull on demand.

The design consequence is a priority rule stated in the README: if the corpus covers a topic, quote it; if it does not, reason with his method and say so in the first sentence. That second path is what keeps the skill from answering everything with a flat refusal. The example in the README on innovative drugs shows the shape: the skill states up front that Zheng Xi has never publicly discussed the sector, then walks his framework through it and flags that company and valuation specifics need verification.

Installing the skill and asking your first sourced question

For Claude Code the README gives a direct copy into the skills directory. On macOS or Linux, run the following from the repository root. It creates the target folder and copies the manifest, README, references and scripts into it.

mkdir -p "$HOME/.claude/skills/zhengxi-views"
cp -R SKILL.md README.md references scripts "$HOME/.claude/skills/zhengxi-views"/

Windows PowerShell does the same with native cmdlets.

$dst = "$HOME\.claude\skills\zhengxi-views"
New-Item -ItemType Directory -Force $dst | Out-Null
Copy-Item -Recurse SKILL.md,README.md,references,scripts $dst

After copying, the README says to fully restart Claude Code, then ask a question such as what Zheng Xi thinks about optical communications. The expected result is a quoted passage with its year and publication, plus a short account of how the view developed, rather than a paraphrase.

Pure corpus search needs no third-party libraries. Only the fetching and scoring scripts do, and requirements.txt pins three of them.

pip install -r requirements.txt   # requests / beautifulsoup4 / lxml

The file lists requests>=2.28, beautifulsoup4>=4.11 and lxml>=4.9. If you only intend to ask sourced questions from the existing corpus, you can skip this step entirely. The README also notes a skill.yml manifest for Tencent WorkBuddy, which the repository says was tested there, and a WORKBUDDY部署.md file describing deployment.

Scoring a fund with the six-dimension card, and what the number means

The scoring path is the part most likely to be misread. scripts/score_fund.py applies a six-dimension card from references/scorecard.md and returns a total, a rating and reasons. The README is explicit that the score measures how much a fund resembles something Zheng Xi would buy, not whether the fund is any good, and the worked example makes the point: an index fund tracking a single traditional consumer sector scores 21 out of 100 and is labelled a poor fit, while the text immediately says a low score does not mean a bad fund, only that the style is close to the opposite of his.

That framing is honest, and it is also the main thing to keep in mind when reading output. The dimensions in the example are sector direction and inflation character, low-ROE elasticity, global view and China's comparative advantage, liquidity, concentration and cycle-splicing, and performance with drawdown confirmation. A passive sector index will fail most of them by construction. If your question is whether a fund is well run, this card will not answer it, and the README does not claim otherwise.

The market-wide comparison path works differently. scripts/fund_lookup.py resolves a fund code against the roughly 27,000-entry list, and scripts/fetch_any_fund.py retrieves data for it so it can be placed side by side with Zheng Xi's funds. The README's comparison of his information-industry fund with a healthcare fund contrasts turnover style and drawdown, and attributes the figures to public quarterly snapshots current to 2026-06.

Where the skill stops being the right tool

Coverage is the first limit, and it is structural. The corpus is one manager's public writing. If you need to compare five managers, or screen a category, the skill has nothing to say about the other four beyond the market-wide fund list, which supplies codes and names rather than views. The all-funds file supports fetching and scoring, not opinion.

The second limit is that absence of evidence is the normal state for most topics. A discretionary manager writes a few pages a quarter, so the corpus will not cover most sectors, most stocks and most current events. The skill handles this by labelling its output as inference from his method, which is the correct behaviour, but it means a large share of interesting questions produce a derivation rather than a quotation. If you want the quotation and nothing else, expect to be told it does not exist fairly often.

The third is freshness and the fetch path. The last push to the repository was on 2026-09-04, the same day as the v1.2.0 release, which the release notes describe as an update for the 2026 interim report and 2026 Q2 holdings. That is recent, but the corpus and the snapshots are only as current as the last update, and the fetch scripts depend on external sources through requests and parsing with beautifulsoup4 and lxml, so a change in the source page layout can break retrieval. The README does not document a fallback when a fetch fails.

Fourth, the licence. The repository shows NOASSERTION and the README badge says MIT. That mismatch is worth resolving before you redistribute anything, and the corpus itself is third-party published material, which the repository does not discuss from a rights standpoint.

Compared with a generic investing-framework skill

The README draws its own comparison against framework or methodology skills, and the distinction is worth stating plainly because it determines which one you want. A generic framework skill packages an abstracted method and applies it to any security you name. Its output is a selection list or a score, and it does not quote anyone.

zhengxi-views inverts the priority. The corpus is the foundation, and both the forward-looking reasoning and the scoring are anchored to it. The stated output types are sourced quotations, the evolution of a view over time, and forward judgements that carry a citation or an explicit label saying they are an inference. The README summarises the constraint as: quote when the corpus has it, mark it as reasoning from his method when it does not.

If your goal is to generate candidates across a universe, the framework skill is the better fit, because this project will keep stopping to ask whether Zheng Xi ever said anything about the name. If your goal is to understand one manager and hold his statements against his filings, the sourced approach is the one that produces something you can check. The trade-off is coverage against verifiability, and this project sits firmly on the verifiability side.

Editorial conclusion

Adopt zhengxi-views if you want sourced answers about Zheng Xi's published views and a scoring pass built on his stated method, and you accept that the output is a research aid rather than advice. Do not adopt it if you need coverage of managers other than Zheng Xi, or if you cannot run the Python scripts that fetch fund data. Before relying on it, verify two things yourself: that the corpus files under references/corpus/ actually contain the passages you intend to cite, and that the fund snapshots in references/fund_data/ match the disclosure periods you need. The scoring script measures resemblance to what Zheng Xi tends to buy, not whether a fund is good, and the README says so.

Frequently asked questions

Official sources

1. Issues
2. lyra81604/zhengxi-views on GitHub
3. Project website
4. README
5. Releases